mirror of https://github.com/OpenIdentityPlatform/OpenDJ.git
opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/JDBCStorage.java
@@ -44,6 +44,7 @@ import java.util.*; import java.util.concurrent.ConcurrentHashMap; import java.util.concurrent.Executor; import java.util.concurrent.TimeUnit; import java.util.concurrent.atomic.AtomicBoolean; import java.util.function.Predicate; @@ -58,15 +59,25 @@ private static final int MAX_RETRIES = 10; /** * Wall-clock budget the replays of a {@link #write} may spend, in nanoseconds. It is checked between attempts, * so an attempt already running is never interrupted: the loop returns after at most this window plus one * attempt. It bounds the conflicts that are slow to report, which {@link #MAX_RETRIES} alone does not - MySQL * reports a lock wait timeout only after innodb_lock_wait_timeout, 50 s by default and not overridden here, so * ten attempts would park a worker thread for eight minutes where a single one released it after 50 s. The * deadlocks this retry exists for keep their full attempt budget, since every engine reports one in well under * a second. * Wall-clock budget the replays of a {@link #write} may spend, in nanoseconds, measured from the start of the * first attempt. It is checked between attempts, so an attempt already running is never interrupted, and it * applies from the first check, with the single exception {@link #grantedPastTheWindow} describes: a conflict * its engine reports promptly is granted one replay whatever the clock says, because the lock wait that * precedes such a conflict is charged to the attempt and is unbounded on three of the four engines here, so no * window survives it. It bounds what {@link #MAX_RETRIES} alone does not - MySQL reports a lock wait timeout * only after innodb_lock_wait_timeout, 50 s by default and not overridden here, so ten attempts would park a * worker thread for eight minutes where one releases it after 50 s. * <p> * What that costs, stated rather than left to be read off a test row: at the stock innodb_lock_wait_timeout a * MySQL lock wait timeout is reported at ~50 s, which is past this window on the first check, so such a write * is never replayed at all - the one conflict class of the set that a MySQL deployment sees most, and the one * whose replay would most reliably succeed. It is the deliberate half of the trade the other half of which is * #903: one bounded wait beats two, and a deployment that tunes innodb_lock_wait_timeout below this window * gets its replays back. The trade only exists because nothing here bounds the attempt: with a session lock * timeout on the transaction connection (#915) every wait would be shorter than this window, the tuned-down * case would become the normal one, and this window would govern both classes with no grant needed at all. */ private static final long MAX_RETRY_WINDOW_NANOS = 10L * 1000L * 1000L * 1000L; //10 s private static final long RETRY_WINDOW_NANOS = TimeUnit.SECONDS.toNanos(10); /** Upper bound of the random delay before the second attempt, in milliseconds; it doubles with every attempt. */ private static final double BASE_SLEEP_ON_RETRY_MS = 50.0; @@ -75,15 +86,30 @@ private static final double MAX_SLEEP_ON_RETRY_MS = 1000.0; /** * Number of links walked when classifying a failure, also a guard against a chain long enough to matter. One * number for three chains at once - the causes, the next exceptions and the suppressed exceptions are walked * together and counted together - so it is set well above the depth a wrapped failure of this backend reaches: * mssql-jdbc chains every error of one message it received through {@code setNextException}, and a budget spent * on those would never reach the cause the wrapper carries. * Number of links walked by the questions that are not asked to the end of the chains: a guard against a chain * long enough to matter, at the cost of what truncation costs each of them. One number for three chains at * once - the causes, the next exceptions and the suppressed exceptions are walked together and counted * together - so it is set well above the depth a wrapped failure of this backend reaches: mssql-jdbc chains * every error of one message it received through {@code setNextException}, and a budget spent on those would * never reach the cause the wrapper carries. * <p> * For the fallbacks of {@link #conflictSummary} truncation leaves a question unanswered and nothing more: the * line reporting the replay names a less precise link, and no decision moves. For * {@link #isConnectionFailure} it does weaken the verdict, which is the test {@link #EVERY_LINK} exists to * apply: a class 08 link past this many links leaves {@code dropped} false in {@link #write}, so * {@link #distrustPool} is not called and the pool keeps handing out - unvalidated - the connections it had * established before the same restart or failover. That is what this walk did before #903 and it is left as it * is here rather than widened along with the two below, since nothing about the grant of #903 depends on it; * the budget is pinned from both sides by {@code testTheWalkOfAFailureStopsAtItsBudget}, which is where * widening it would have to start. */ private static final int MAX_CHAIN_LINKS = 64; /** The budget of {@link #failureScope}, which walks to the end of the chains: see the comment above it. */ /** * The budget of the two walks whose verdict would weaken rather than go unnoticed under truncation: * {@link #failureScope} and {@link #conflictVerdict}. See the comments above them; the {@code seen} set of * the walk terminates it either way. */ private static final int EVERY_LINK = Integer.MAX_VALUE; /** SQL Server error number of the transaction picked as the deadlock victim: "Rerun the transaction". */ @@ -93,6 +119,15 @@ private static final int ORACLE_DEADLOCK_DETECTED = 60; /** * MySQL error number of a lock wait timeout, ER_LOCK_WAIT_TIMEOUT: the one conflict of the set an engine * reports late rather than promptly. Connector/J maps it to the same class 40 state as a deadlock, so this * number is not what matches it - it only tells the two apart, and only under a MySQL driver. That it repeats * the literal of {@link #MSSQL_DEADLOCK_VICTIM} is the collision {@link #conflictVerdict} keys every vendor * number by the driver for, and is stated there rather than again here. */ private static final int MYSQL_LOCK_WAIT_TIMEOUT = 1205; /** * Class 40 states that are transaction rollbacks but must not be replayed. 40003 leaves the outcome of the * transaction unknown, so replaying an add that in fact committed would answer the client with * "entry already exists", and 40002 is an integrity constraint violation, which a replay repeats rather than @@ -537,6 +572,11 @@ // below turns on a few milliseconds either side of the bound, and a mock statement cannot be made // to take a real second without the suite taking one too. Monotonic, so that a step of the wall // clock can neither lengthen nor shorten what a statement is measured to have taken. // // The retry window of write() is read off this same clock, once before the first attempt and once // after each: what that window bounds is the whole run of attempts rather than each attempt on its // own, which is a single startedAt outside the loop. #877 and #903 each added a clock of their own // here, for the same reason and with the same body; this is the one they share. long nanoTime() { return System.nanoTime(); } @@ -1267,14 +1307,26 @@ // The dialect behind a pooled connection, or null for an engine none of the statements of this // class fit: it is left unstamped and its statistics untouched rather than fed untested SQL. static Dialect dialectOf(Connection con) { final String driverName=driverNameOf(con); if (driverName.contains("postgres")) { return dialectOf(driverNameOf(con)); } /** * The engine a driver class name names, or null for one this class does not recognise. Every question this * class asks about the engine is keyed on this one answer - the column types, the upsert, the paging clause, * whether a DDL statement commits, and the class of a conflict - so that they cannot disagree about a driver. * A cascade of its own per question is how a deployment ends up given one engine's SQL and another engine's * conflict class; what an unrecognised engine gets is a decision per question, taken and stated at each of * them rather than falling out of the order the {@code contains} calls happen to be written in. */ static Dialect dialectOf(String driverName) { final String name=String.valueOf(driverName); if (name.contains("postgres")) { return Dialect.POSTGRES; }else if (driverName.contains("mysql")) { }else if (name.contains("mysql")) { return Dialect.MYSQL; }else if (driverName.contains("oracle")) { }else if (name.contains("oracle")) { return Dialect.ORACLE; }else if (driverName.contains("microsoft")) { }else if (name.contains("microsoft")) { return Dialect.MICROSOFT; } return null; @@ -1654,9 +1706,12 @@ private WriteableTransactionTransactionImpl txn; // The moment the replay window of the write() this session belongs to runs out, as // System.nanoTime() reads it - null where nothing above this session replays, which is the // importer and nothing else. Boxed rather than given a sentinel: nanoTime() is documented to // return an arbitrary long, so there is no reading of it that could stand for "no window". // JDBCStorage.nanoTime() reads it - null where nothing above this session replays, which is the // importer and nothing else. That clock and not System.nanoTime() directly: the window this is a // reading of was taken from it, and the two are the same clock everywhere except the one place // they would be compared as a mixed pair. Boxed rather than given a sentinel: nanoTime() is // documented to return an arbitrary long, so there is no reading of it that could stand for // "no window". private Long replayWindowEndsAt; /** @@ -1683,7 +1738,7 @@ if (replayWindowEndsAt==null) { return Long.MAX_VALUE; } final long leftNanos=replayWindowEndsAt-System.nanoTime(); final long leftNanos=replayWindowEndsAt-nanoTime(); final long now=System.currentTimeMillis(); return leftNanos<=0 ? now : now+leftNanos/1_000_000L; } @@ -2944,10 +2999,15 @@ * on the conflict exception of its own engine. The loop is bounded here, unlike PDBStorage: the database may be * shared with writers outside this server, so a conflict is not guaranteed to clear and failing the operation is * better than never returning. It is bounded twice - by {@link #MAX_RETRIES} attempts and by the * {@link #MAX_RETRY_WINDOW_NANOS} wall-clock window - because an attempt is not guaranteed to be short: a * conflict an engine reports only after its own lock wait timeout would otherwise multiply that wait by the * attempt count. A conflict that slow consumes the whole window in one attempt and is not replayed, which is * what master did with it. * {@link #RETRY_WINDOW_NANOS} wall-clock window - because an attempt is not guaranteed to be short: a conflict * an engine reports only after its own lock wait timeout would otherwise multiply that wait by the attempt * count. The window alone is not enough either, in the other direction: it is shorter than the wait that * precedes a conflict the engine reports promptly, so measured against such a conflict it does not bound that * wait but only leaves the operation with no replay at all, which is what master did with the deadlock of * issue #903. One replay is therefore granted to a prompt conflict whatever the clock says; see * {@link #grantedPastTheWindow}. The window is checked between attempts, so an attempt already running is * never interrupted: a conflicted operation holds its caller for the window plus one attempt, and a prompt * conflict for two attempts when that is longer. * <p> * Only the operation itself is replayed: a failure of {@link #getConnection()} or of the implicit * {@link Connection#close()} - which returns the connection to the pool after a rollback - leaves the loop, so @@ -2958,11 +3018,11 @@ * connection handed out unvalidated and found dead costs an attempt rather than the operation, and a write of * the replication replay - which records a failed operation as applied and advances the server state past it, * see #889 - never sees it. Only while nothing of the attempt may have been committed yet, though: see * {@link #replayReason(Throwable, String, boolean, boolean, boolean)}. * {@link #replayReason(Conflict, Throwable, boolean, boolean, boolean)}. */ @Override public void write(WriteOperation writeOperation) throws Exception { final long giveUpAt=System.nanoTime()+MAX_RETRY_WINDOW_NANOS; final long startedAt=nanoTime(); for (int attempt=1;;attempt++) { Exception failure=null; String driver=null; @@ -2978,8 +3038,12 @@ //the connect of the catalog is made inside this attempt and retries the way a borrow does, //up to the pool timeout - six times this window at the defaults. Left to its own deadline //it would spend a window it does not own and hand the loop a failure it has classified as //replayable with nothing left to replay it in, so it is told where the window ends txn.catalogSession.boundedAlsoBy(giveUpAt); //replayable with nothing left to replay it in, so it is told where the window ends. The //end of the window and not what is left of it: this is called once per attempt, and a //window measured from the attempt that happens to be running would be spent over again by //each of them. The grant of #903 is deliberately not passed on - it buys one more replay //of the operation, not one more connect of the catalog inside it txn.catalogSession.boundedAlsoBy(startedAt+RETRY_WINDOW_NANOS); try { writeOperation.run(txn); committing=true; @@ -3055,15 +3119,42 @@ if (!dropped && isConnectionFailure(failure)) { distrustPool(); } final String reason=replayReason(failure,driver,committing,partlyCommitted,dropped); //System.nanoTime()-giveUpAt is the overflow safe form of the comparison if (reason==null || attempt>=MAX_RETRIES || System.nanoTime()-giveUpAt>=0) { //Two questions, asked apart: what the failure is - which replayReason() answers, and which is the //only place that reads committing, partlyCommitted and dropped - and whether another attempt is //still allowed, which is the attempt count and the window of #903. Neither subsumes the other: a //dropped connection is worth replaying and carries no conflict class, while a conflict past both //bounds is not replayed however plainly it is one //classified once and handed to every question below - the two decisions and the line reporting //them: the walk of the chains is not free, and callers asking it apart could drift into //disagreeing about the same failure. Not asked at all where the answer is discarded: replayReason() //refuses a partly committed attempt its replay before anything about the failure matters, and that //is also the path most likely to carry deeply wrapped chains, since RootContainer.open() commits //DDL and raises the flag for the rest of the write final ConflictVerdict verdict=partlyCommitted ? NOT_CLASSIFIED : conflictVerdict(failure,driver); final String reason=replayReason(verdict.conflict,failure,committing,partlyCommitted,dropped); //nanoTime()-startedAt is the overflow safe form of the elapsed time final long elapsedNanos=nanoTime()-startedAt; if (reason==null || !replayableWithin(attempt, elapsedNanos, verdict.conflict)) { throw failure; } //logged rather than silently absorbed, so that a deployment retrying most of its writes stays observable; //one line per replay, since an add can emit nine of them and a stack trace each time reads as a failure logger.warn(LocalizableMessage.raw("jdbc: replaying the transaction after %s, attempt %d of %d: %s", reason, attempt, MAX_RETRIES, conflictSummary(failure, driver))); //one line per replay, since an add can emit nine of them and a stack trace each time reads as a failure. //Both bounds are named, and the attempt count is the one that rarely fires: a replay usually stops //because the window ran out, and a log naming only MAX_RETRIES leaves an operation that gave up at //attempt 2 of a promised 10 with nothing saying why. Milliseconds rather than seconds, since the //engines report a deadlock in a few of them and whole seconds would read "0" for most of a burst; and //the one line that replays past its own window says so, rather than reading as a bound not honoured - //asked of the predicate the loop just acted on rather than re-derived from the clock, so that the //claim cannot outlive the grant that justifies it if (logger.isWarnEnabled()) { logger.warn(LocalizableMessage.raw( "jdbc: replaying the transaction after %s, attempt %d of %d, %d ms elapsed of the %d ms window%s: %s", reason, attempt, MAX_RETRIES, TimeUnit.NANOSECONDS.toMillis(elapsedNanos), TimeUnit.NANOSECONDS.toMillis(RETRY_WINDOW_NANOS), grantedPastTheWindow(attempt, elapsedNanos, verdict.conflict) ? " (the first replay, granted past it)" : "", conflictSummary(verdict, failure))); } if (logger.isTraceEnabled()) { logger.trace("jdbc: the failure being replayed was %s", stackTraceToSingleLineString(failure)); } @@ -3086,9 +3177,9 @@ * <p> * A transaction conflict is replayable whichever phase reported it: the engine rolled the transaction back * before it answered. It is read from the failure of the operation only, never from the release of the * connection - see {@link #isRetryableConflict} - since the release runs after the outcome was decided and * cannot make that claim for it. A connection the database dropped is replayable only while the transaction * had not been committed yet. A drop reported by {@code commit()} leaves the outcome unknown - the server may * connection - see {@link #conflictVerdict} - since the release runs after the outcome was decided and cannot * make that claim for it. A connection the database dropped is replayable only while the transaction had not been * committed yet. A drop reported by {@code commit()} leaves the outcome unknown - the server may * have committed and died before the answer reached us - and replaying a write that in fact committed applies * it twice, which is the very reason 40003 is one of {@link #NON_REPLAYABLE_ROLLBACK_STATES}. * <p> @@ -3101,17 +3192,18 @@ * second time and fails with ERR_ENTRY_CONTAINER_ALREADY_REGISTERED, which masks the failure that caused the * replay and leaves the indexes of the previous attempt behind with their configuration listeners. * * @param conflict the class {@link #conflictVerdict} read from the failure, asked of it once by the caller * @param committing whether the failure was reported by {@code commit()}, which leaves the outcome unknown * @param partlyCommitted whether the attempt committed part of its work before it failed * @param connectionClosed whether the driver closed the connection under the failure - evidence no SQLState * carries on mssql-jdbc, which reports a killed session as S0001 and closes the connection behind it */ static String replayReason(Throwable failure, String driver, boolean committing, boolean partlyCommitted, static String replayReason(Conflict conflict, Throwable failure, boolean committing, boolean partlyCommitted, boolean connectionClosed) { if (partlyCommitted) { return null; } if (isRetryableConflict(failure, driver)) { if (conflict!=Conflict.NONE) { return "a conflict"; } if (!committing && (connectionClosed || isConnectionFailure(failure))) { @@ -3184,6 +3276,27 @@ /** The walk above, with the number of links it is allowed to look at. */ private static SQLException firstLinkMatching(Throwable failure, boolean withTheRelease, int links, Predicate<SQLException> matches) { final SQLException[] found=new SQLException[1]; walkLinks(failure, withTheRelease, links, e -> { if (!matches.test(e)) { return false; } found[0]=e; return true; }); return found[0]; } /** * Hands every {@link SQLException} of the chains of a failure to the given reader, in walk order, until it * says it has read enough. The single traversal of this class: a reader that can answer from the first link * it matches stops here, and one that has to see them all - {@link #conflictVerdict}, which keeps the * strongest class any of them carries - does not, so that neither has a walk of its own to drift from the * other's. The {@code seen} set terminates the walk whatever budget it is given: a driver that chains an * exception back to itself is walked once. */ private static void walkLinks(Throwable failure, boolean withTheRelease, int links, Predicate<SQLException> readEnough) { final Deque<Throwable> pending=new ArrayDeque<>(); final Set<Throwable> seen=Collections.newSetFromMap(new IdentityHashMap<Throwable,Boolean>()); if (failure!=null) { @@ -3209,11 +3322,10 @@ if (sqlException.getNextException()!=null) { pending.push(sqlException.getNextException()); } if (matches.test(sqlException)) { return sqlException; if (readEnough.test(sqlException)) { return; } } return null; } /** @@ -3251,13 +3363,84 @@ } /** * Returns whether the given failure carries a transaction conflict that replaying the operation can resolve. * The class of a failure. Whether it is a conflict at all is what decides that the operation is replayed; * which of the remaining classes it is decides only whether the first replay is granted unconditionally, * since the wait an engine spends before reporting a conflict is charged to the attempt that hit it. * <p> * Declared in order of how much they restrict the replay, which is the order {@link #conflictVerdict} * compares them in: the strongest class any link of a failure carries is the class of that failure. Only * {@link #PROMPT} is granted the replay past the window, so every class this list gains - #915 adds one - * has to be placed against that grant rather than merely appended. */ enum Conflict { /** Not a conflict: no replay resolves it. */ NONE, /** A conflict reported as soon as the engine detects it, however long the attempt waited to reach it. */ PROMPT, /** * A conflict under a driver none of the four engines is recognised in. Whether the engine bounded the * wait that preceded it is not something this class can tell, and the grant of {@link #PROMPT} rests on * knowing that it did not, so an unrecognised engine is refused it: see {@link #classOf}. */ UNKNOWN_ENGINE, /** A conflict an engine reports only once a lock wait timeout of its own has elapsed. */ AFTER_LOCK_WAIT } /** * The verdict of a failure nothing asked about, handed to the questions {@link #write} asks of a partly * committed attempt: every one of them is answered by that flag alone, so its chains are never walked. It is * not a claim that the failure carries no conflict - it may carry one, and is refused a replay either way. */ private static final ConflictVerdict NOT_CLASSIFIED=new ConflictVerdict(Conflict.NONE, null); /** * The class of the conflict a failure carries and the link that class was read from, which are one answer * rather than two: the line reporting a replay names the link the decision was taken on, and a summary that * walked the chains again to find it could name a different one - see {@link #conflictSummary}. */ static final class ConflictVerdict { final Conflict conflict; /** Null where the failure carries no conflict at all, which is what {@link Conflict#NONE} says. */ final SQLException link; ConflictVerdict(Conflict conflict, SQLException link) { this.conflict=conflict; this.link=link; } } /** * Returns the class of the conflict the given failure carries, or {@link Conflict#NONE} if it carries none - * which is what decides whether replaying the operation can resolve it - together with the link that class was * read from. One walk of the chains that keeps the strongest class it meets, rather than one walk per class * asked in the right order: asking per class is what let the two walks of the earlier form be given different * budgets, and the ordering of an added class is then a rule its author has to find rather than one the enum * states - see {@link Conflict}. * <p> * The conflict is looked up along every chain of the failure, for the reason {@link #isConnectionFailure} walks * them all: it reaches this class wrapped - a deadlock in {@code put} arrives as * {@code StorageRuntimeException(SQLException)}, and a caller such as {@code EntryContainer.addEntry} may wrap it * once more - and a driver reports the error that says what happened as the next exception of a generic one at * least as often as it reports it as the cause. * least as often as it reports it as the cause. The suppressed links of the release are left out of it, for the * reason {@link #replayReason} gives: the release runs after the outcome was decided. * <p> * The strongest class in those chains wins rather than the first one found: a wrapper that carries a class * 40 state of its own but no vendor number would otherwise downgrade the {@link Conflict#AFTER_LOCK_WAIT} of * the {@link SQLException} it wraps, and hand a wait the engine already bounded a replay it does not need. * That rule is deliberately not restricted to the wrapper it was introduced for, although the walk reaches * links that are not ancestors of the operative failure - a deadlock whose chain also carries a lock wait * timeout is classed by the timeout and loses the grant. The two errors are not equally costly to get wrong: * granting a replay to a wait the engine had already bounded pays that bound a second time, while refusing * one to a deadlock costs a replay the window was about to refuse anyway, wherever the bound that sibling * names is longer than the window. So the class is read the conservative way, and the whole chain of a * failure is evidence for it. * <p> * Every link is looked at, rather than {@link #MAX_CHAIN_LINKS} of them, for the reason {@link #failureScope} * walks to the end: the verdict weakens under truncation rather than simply going unnoticed. An * {@link Conflict#AFTER_LOCK_WAIT} link past the budget with a bare class 40 link inside it comes back * {@link Conflict#PROMPT}, and truncation there does not lose a replay - it grants one, which is the single * thing this classification exists to refuse. The {@code seen} set terminates the walk regardless. * <p> * The standard class 40 states carry the conflict of most engines - 40P01 for PostgreSQL, 40001 for SQL Server * and for MySQL, whose driver replaces the server side HY000 of a deadlock and of a lock wait timeout with @@ -3266,28 +3449,135 @@ * reports a deadlock as ORA-00060 with SQLState 61000, and gives 1205 to a fatal "not a data file" error that * no replay can resolve, while 1205 is exactly the deadlock victim of SQL Server. The SQL Server number is * matched beyond its class 40 state because a deployment may add {@code xopenStates=true} to its connection * URL, which reports the same deadlock as 42000. MySQL needs no number of its own, since its driver has already * mapped both conditions into class 40; see {@link #NON_REPLAYABLE_ROLLBACK_STATES} for the two class 40 states * that are excluded from that match. * URL, which reports the same deadlock as 42000. MySQL needs no number of its own for the match, since its * driver has already mapped both conditions into class 40 - its number is read by {@link #classOf} alone, and * only to tell the two apart; see {@link #NON_REPLAYABLE_ROLLBACK_STATES} for the two class 40 states that are * excluded from that match. * <p> * The walk still stops as soon as its answer is final, but the class it stops at is the strongest one this * engine can report - {@link #ceilingOf} - rather than the strongest one the enum declares. Only MySQL reports * an {@link Conflict#AFTER_LOCK_WAIT}, so a walk stopping at that constant never stops early on the other * three engines, nor under a driver none of them is recognised in: it reads every link of every failed write, * a plain {@code 23000} from adding an entry that is already there included, on the driver whose chains are * longest. The dialect is resolved once here for the same reason - {@link #classOf} and {@link #isConflict} * would otherwise read it off the driver name twice for every link walked. */ static boolean isRetryableConflict(Throwable t, String driver) { // without the suppressed exceptions, unlike isConnectionFailure(): a conflict is replayed whichever phase // reported it, on the strength of the engine having rolled the transaction back before it answered - and // the release of the connection runs after the outcome was decided and cannot make that claim. A class 40 // raised there would otherwise replay a transaction commit() left in doubt, which is what the committing // guard of replayReason() exists to prevent return firstLinkMatching(t, WITHOUT_THE_RELEASE, e -> isConflict(e, driver))!=null; static ConflictVerdict conflictVerdict(Throwable failure, String driver) { final Dialect dialect=dialectOf(driver); final Conflict ceiling=ceilingOf(dialect); final Conflict[] strongest={Conflict.NONE}; final SQLException[] link=new SQLException[1]; walkLinks(failure, WITHOUT_THE_RELEASE, EVERY_LINK, e -> { final Conflict conflict=classOf(e, dialect); if (conflict.compareTo(strongest[0])>0) { strongest[0]=conflict; link[0]=e; } return strongest[0]==ceiling; }); return new ConflictVerdict(strongest[0], link[0]); } private static boolean isConflict(SQLException e, String driver) { /** * The strongest class a conflict raised under the given engine can carry, which is where * {@link #conflictVerdict} stops walking: nothing further along the chains can outrank it. It is the maximum * of what {@link #classOf} returns for that dialect and has to be read together with it - a property of the * engine rather than the last constant of {@link Conflict}, so that the class #915 adds cannot silently move * the stop condition, and so that the walk of the three engines reporting no lock wait timeout of their own * ends on the first conflict it meets rather than at the end of every chain. */ static Conflict ceilingOf(Dialect dialect) { if (dialect==null) { return Conflict.UNKNOWN_ENGINE; } return dialect==Dialect.MYSQL ? Conflict.AFTER_LOCK_WAIT : Conflict.PROMPT; } /** * Returns the class of a single failure. The vendor number only refines a failure {@link #isConflict} has * already matched and never widens that match, which the engines colliding on 1205 do not allow: the number is * read here to tell the late conflict of MySQL from the deadlock its driver reports under the same state. * <p> * A conflict raised under a driver {@link #dialectOf(String)} did not recognise is * {@link Conflict#UNKNOWN_ENGINE}: still replayed, since replayability is what the class 40 state says and it * says it whatever the engine, but not granted the replay past the window. The grant rests on knowing that * the wait preceding the conflict was not bounded by the engine, and of an unrecognised engine that is not * known. It is a MySQL-wire-compatible driver - MariaDB Connector/J, an Aurora- or Percona-branded one - * that makes the difference concrete: it reports a lock wait timeout as 1205 under class 40 exactly as * Connector/J does, this class would read the number only under a name carrying {@code mysql}, and granting * a free replay there buys a second full {@code innodb_lock_wait_timeout}. Such a deployment does reach this * code: a backend created under {@code com.mysql.cj.jdbc} and later opened through one of those drivers * issues no DDL at all - every {@code create table} and {@code create index} of * {@code openTree(createOnDemand)} is guarded by a catalog read - and its writes go down the ANSI branch of * {@code upsert}, which is an {@code update} and an {@code insert}, not a statement a MySQL-wire engine * refuses. The cost of the class is one replay of the window's own length for an engine whose conflicts are * in fact prompt, which is the direction worth being wrong in; #915 removes the trade by bounding the * attempt itself. */ private static Conflict classOf(SQLException e, Dialect dialect) { if (!isConflict(e, dialect)) { return Conflict.NONE; } if (dialect==null) { return Conflict.UNKNOWN_ENGINE; } return dialect==Dialect.MYSQL && e.getErrorCode()==MYSQL_LOCK_WAIT_TIMEOUT ? Conflict.AFTER_LOCK_WAIT : Conflict.PROMPT; } /** * Whether another attempt is still allowed: the bounds half of the decision {@link #write} takes after every * attempt, asked of a failure {@link #replayReason} has already found worth replaying and made here apart * from the clock so that it can be tested without a database. It is asked of the conflict class rather than * of the failure because not every replayable failure carries one - a connection the database dropped is * replayed on the evidence of the drop, and would be refused by a bound that first insisted on a class 40 * state - and because {@code write()} has already read that class off the failure once. * <p> * Replays are bounded by {@link #MAX_RETRIES} and by {@link #RETRY_WINDOW_NANOS} against the time elapsed * since the first attempt began, with the one grant {@link #grantedPastTheWindow} states on top of them. */ static boolean replayableWithin(int attempt, long elapsedNanos, Conflict conflict) { if (attempt>=MAX_RETRIES) { return false; } if (grantedPastTheWindow(attempt, elapsedNanos, conflict)) { return true; } return elapsedNanos<RETRY_WINDOW_NANOS; } /** * Whether this replay is the one {@link #RETRY_WINDOW_NANOS} does not get to deny: the first replay of a * conflict its engine reports promptly, taken although the window is already spent. The wait an engine spends * before reporting such a conflict is charged to the attempt that hit it and is unbounded on three of the four * engines here - SQL Server took some 12 s to pick a victim in CI - so there is no window that some wait does * not outlast, and measuring one against it only leaves the operation with no replay at all, which is issue * #903. The grant does not extend to {@link Conflict#AFTER_LOCK_WAIT}, whose wait the engine has already * bounded for us: replaying that costs the same bounded wait again, which is exactly what the window is here * to refuse. Nor to {@link Conflict#UNKNOWN_ENGINE}, of which the same cannot be ruled out. * <p> * Asked as a question of its own so that the line reporting the replay can name the bound that was actually * applied instead of inferring it from the clock: {@code elapsed >= window} coincides with this grant only * for as long as this stays the sole way past the window, and a line that keeps claiming "the first replay" * after that would be describing a decision nobody took. * <p> * {@code attempt==1} is a proxy and not the invariant: the invariant is that no clock can bound a wait * nothing else bounds, and that holds on every attempt, not only the first. Widening the grant to all of them * would leave {@link #MAX_RETRIES} as the only real cap, so it is held to one replay until the attempt itself * carries a lock bound - see #915, which retires this method rather than widening it. */ static boolean grantedPastTheWindow(int attempt, long elapsedNanos, Conflict conflict) { return attempt==1 && conflict==Conflict.PROMPT && elapsedNanos>=RETRY_WINDOW_NANOS; } private static boolean isConflict(SQLException e, Dialect dialect) { final String state=String.valueOf(e.getSQLState()); if (state.startsWith("40") && !NON_REPLAYABLE_ROLLBACK_STATES.contains(state)) { return true; } final String driverName=String.valueOf(driver); if (driverName.contains("oracle")) { if (dialect==Dialect.ORACLE) { return e.getErrorCode()==ORACLE_DEADLOCK_DETECTED; } else if (driverName.contains("microsoft")) { } else if (dialect==Dialect.MICROSOFT) { return e.getErrorCode()==MSSQL_DEADLOCK_VICTIM; } return false; @@ -3301,11 +3591,16 @@ * suppressed into it, and naming the state of the rejected statement instead would describe a replay that did * not happen. Falls back to the first SQLException of the failure, and to the failure itself where it carries * none. * <p> * Asked of the verdict rather than of the failure and the driver, since {@link #write} - the only caller - has * had the failure classified already: a form taking those two would walk the chains a second time to reach the * verdict this one is handed. */ static String conflictSummary(Throwable failure, String driver) { // asked in the order replayReason() asks it, and of the same chains, so that the line names the link the // decision was taken on rather than one that merely resembles it SQLException named=firstLinkMatching(failure, WITHOUT_THE_RELEASE, e -> isConflict(e, driver)); static String conflictSummary(ConflictVerdict verdict, Throwable failure) { // the link the class was read from, handed over by the walk that read it rather than looked up again in // the order that walk happens to use: repeated by hand, the two drift, and the line then names a link // that merely resembles the one the decision was taken on SQLException named=verdict.link; if (named==null) { named=firstLinkMatching(failure, WITH_THE_RELEASE, JDBCStorage::saysTheConnectionIsGone); } @@ -3359,7 +3654,7 @@ * Casting the parameter back to char keeps the comparison seekable. */ static String hashParam(Connection con) { return driverNameOf(con).contains("microsoft") ? "cast(? as char(128))" : "?"; return dialectOf(con)==Dialect.MICROSOFT ? "cast(? as char(128))" : "?"; } class ReadableTransactionImpl implements ReadableTransaction { @@ -3570,19 +3865,20 @@ /** Whether this engine commits the transaction before a DDL statement whether asked to or not. */ private boolean commitsBeforeDdl() { final String driverName=driverNameOf(con); return driverName.contains("mysql") || driverName.contains("oracle"); final Dialect dialect=dialectOf(con); return dialect==Dialect.MYSQL || dialect==Dialect.ORACLE; } String getTableDialect() { if (driverNameOf(con).contains("oracle")) { final Dialect dialect=dialectOf(con); if (dialect==Dialect.ORACLE) { return "h char(128),k raw(2000),v blob,primary key(h,k)"; }else if (driverNameOf(con).contains("mysql")) { }else if (dialect==Dialect.MYSQL) { return "h char(128),k varbinary(255),v longblob,primary key(h,k)"; }else if (driverNameOf(con).contains("microsoft")) { }else if (dialect==Dialect.MICROSOFT) { return "h char(128),k varbinary(max),v image,primary key(h)"; } return "h char(128),k bytea,v bytea,primary key(h,k)"; return "h char(128),k bytea,v bytea,primary key(h,k)"; // postgres, and an unrecognised engine with it } @Override @@ -3616,9 +3912,9 @@ } } // CursorImpl iterates with "where k>? order by k" batches: primary key (h,k) cannot serve them final String driverName=driverNameOf(con); final Dialect dialect=dialectOf(con); final String tableName=getTableName(treeName); if (driverName.contains("postgres")) { if (dialect==Dialect.POSTGRES) { try { // asked although postgresql has "create index if not exists": that statement commits // whether it creates anything or not, and this is the engine of every default @@ -3630,7 +3926,7 @@ }catch (SQLException e) { throw new StorageRuntimeException(e); } }else if (driverName.contains("mysql")) { }else if (dialect==Dialect.MYSQL) { try { if (!isExistsIndex(tableName,"k_"+tableName.substring("opendj_".length()))) { // mysql has no "create index if not exists" commitStatement("create index k_"+tableName.substring("opendj_".length())+" on "+tableName+" (k)", true); @@ -3638,7 +3934,7 @@ }catch (SQLException e) { throw new StorageRuntimeException(e); } }else if (driverName.contains("oracle")) { }else if (dialect==Dialect.ORACLE) { try { // oracle has no "create index if not exists"; unquoted identifiers are stored in uppercase if (!isExistsIndex(tableName.toUpperCase(Locale.ROOT),"k_"+tableName.substring("opendj_".length()))) { @@ -4065,29 +4361,29 @@ } boolean upsert(TreeName treeName, ByteSequence key, ByteSequence value) throws SQLException { final String driverName=driverNameOf(con); if (driverName.contains("postgres")) { //postgres upsert final Dialect dialect=dialectOf(con); if (dialect==Dialect.POSTGRES) { //postgres upsert try (final PreparedStatement statement = con.prepareStatement("insert into " + getTableName(treeName) + " (h,k,v) values (?,?,?) ON CONFLICT (h, k) DO UPDATE set v=excluded.v")) { statement.setString(1, key2hash.get(ByteBuffer.wrap(key.toByteArray()))); statement.setBytes(2, real2db(key.toByteArray())); statement.setBytes(3, value.toByteArray()); return (execute(statement, bound) == 1 && statement.getUpdateCount() > 0); } }else if (driverName.contains("mysql")) { //mysql upsert }else if (dialect==Dialect.MYSQL) { //mysql upsert try (final PreparedStatement statement = con.prepareStatement("insert into " + getTableName(treeName) + " (h,k,v) values (?,?,?) as new ON DUPLICATE KEY UPDATE v=new.v")) { statement.setString(1, key2hash.get(ByteBuffer.wrap(key.toByteArray()))); statement.setBytes(2, real2db(key.toByteArray())); statement.setBytes(3, value.toByteArray()); return (execute(statement, bound) == 1 && statement.getUpdateCount() > 0); } }else if (driverName.contains("oracle")) { //ANSI MERGE without ; }else if (dialect==Dialect.ORACLE) { //ANSI MERGE without ; try (final PreparedStatement statement = con.prepareStatement("merge into " + getTableName(treeName) + " old using (select ? h,? k,? v from dual) new on (old.h=new.h and old.k=new.k) WHEN MATCHED THEN UPDATE SET old.v=new.v WHEN NOT MATCHED THEN INSERT (h,k,v) VALUES (new.h,new.k,new.v)")) { statement.setString(1, key2hash.get(ByteBuffer.wrap(key.toByteArray()))); statement.setBytes(2, real2db(key.toByteArray())); statement.setBytes(3, value.toByteArray()); return (execute(statement, bound) == 1 && statement.getUpdateCount() > 0); } }else if (driverName.contains("microsoft")) { //ANSI MERGE with ; WITH (HOLDLOCK) makes the upsert atomic: without it SQL Server MERGE can race two concurrent NOT MATCHED inserts of the same key into a PRIMARY KEY violation. UPDLOCK is required on top of it: with HOLDLOCK alone the search phase takes a shared lock that the WHEN MATCHED update then has to convert to an exclusive one, so two concurrent upserts of the same key deadlock on the conversion; an update lock is taken right away and makes the second transaction wait instead. h is cast back to char so that the join can seek the primary key instead of scanning the whole table under those locks, see hashParam() }else if (dialect==Dialect.MICROSOFT) { //ANSI MERGE with ; WITH (HOLDLOCK) makes the upsert atomic: without it SQL Server MERGE can race two concurrent NOT MATCHED inserts of the same key into a PRIMARY KEY violation. UPDLOCK is required on top of it: with HOLDLOCK alone the search phase takes a shared lock that the WHEN MATCHED update then has to convert to an exclusive one, so two concurrent upserts of the same key deadlock on the conversion; an update lock is taken right away and makes the second transaction wait instead. h is cast back to char so that the join can seek the primary key instead of scanning the whole table under those locks, see hashParam() try (final PreparedStatement statement = con.prepareStatement("merge into " + getTableName(treeName) + " WITH (HOLDLOCK, UPDLOCK) old using (select cast(? as char(128)) h,? k,? v) new on (old.h=new.h and old.k=new.k) WHEN MATCHED THEN UPDATE SET old.v=new.v WHEN NOT MATCHED THEN INSERT (h,k,v) VALUES (new.h,new.k,new.v);")) { statement.setString(1, key2hash.get(ByteBuffer.wrap(key.toByteArray()))); statement.setBytes(2, real2db(key.toByteArray())); @@ -4204,7 +4500,7 @@ // of #873 reads the shared tree, and reading a tree must not put it up for removal this.tableName=readTableName(treeName); this.batchBound=batchBound; this.limitClause=((CachedConnection)con).parent.getClass().getName().contains("mysql") this.limitClause=dialectOf(con)==Dialect.MYSQL ? " limit ?,?" : " offset ? rows fetch next ? rows only"; }
